A Dynamic Online Traffic Classification Methodology Based on Data Stream Mining

Recently, traffic classification (TC) becomes more and more important for network management and measurement tasks. The new-coming machine learning based classification methods can achieve high classification accuracy and fast identification ability; however, all these related TC methods up to now always have the assumption of the stability of classification model constituted from network traffic. It is not true since seldom real-world traffic is static. In this paper, we make a first step towards classifying dynamic online traffic in a data stream perspective to handle the dynamic real-time network traffic. In this paper, we validate the dynamic feature of real-world traffic for the first time, using concept drift from two different levels: overall traffic level and application level. The conclusion convinces us that the user behavior reflected in traffic can vary dramatically due to different conditions and different periods. We then propose a novel integrated dynamic online traffic classification framework; called DSTC (data stream based traffic classification). This DSTC differs from previous work since it aims to deal with dynamic traffic with online identification ability. It is a more realistic framework in which training phase can go simultaneously with classification phase and more accurate training model can be constructed with the feedback from classification result. Experiment results have shown that DSTC can have a high stable classification accuracy of above 95% for network traffic with different periods and user conditions, while accuracy for the traditional classification methodology can vary from 81% to 97% when dealing with different traffic.